OpenMed-PPSN-v4

Model Summary

This model extends OpenMed/OpenMed-PII-SuperClinical-Large-434M-v1 with Irish PPSN detection (B-PPSN / I-PPSN) while preserving baseline behavior on non-PPSN entities.

Intended Use

• Clinical and administrative de-identification pipelines.
• Hybrid usage is recommended: model detection + deterministic PPSN checksum validation.

Out-of-Scope

• Not a legal/compliance determination system.
• Not validated for every institution, note format, OCR stream, or jurisdiction.

Training Data

• Synthetic PPSN span-labeled data generated with the official checksum rules.
• Additional benchmarking reference: nvidia/Nemotron-PII (CC-BY-4.0).

Evaluation Snapshot

• Synthetic non-PPSN F1 (base): 0.8701
• Synthetic non-PPSN F1 (candidate): 0.8420
• Candidate-vs-base agreement F1 on real-text proxy set: 0.9814
• Entities/1k chars (base): 11.4124
• Entities/1k chars (candidate): 11.1231

Usage

Load with transformers token classification pipeline, or integrate with a PPSN checksum validator in your masking service.

Limitations

• Real-text drift check used a proxy general-text set.
• Production deployment requires local validation on your own data.

Licensing and Attribution

• Base model is tagged Apache-2.0.
• This release is Apache-2.0.
• Dataset attribution for CC-BY-4.0 sources is preserved in NOTICE.